Turn resolution into megapixels and a budget
Compute cost for AI image generation scales with resolution, not just the number of images — a 2048×2048 render takes roughly four times the GPU work of a 1024×1024 one. Before you kick off a batch of hundreds or thousands of images, this estimator converts your dimensions into megapixels and maps that onto an approximate credit or USD cost for Stability AI, Replicate, RunPod or OpenAI Images.
How the estimate is calculated
Megapixels are simply the total pixel count divided by a million:
megapixels_per_image = (width × height) / 1,000,000
total_megapixels = megapixels_per_image × image_count
estimated_cost = total_megapixels × platform_rate_per_MPEach platform carries its own per-megapixel rate. Credit-based services like Stability AI charge in abstract credits; GPU-second services like Replicate and RunPod translate compute time into dollars; and per-image APIs like OpenAI Images are mapped to an approximate per-megapixel equivalent so they sit on the same scale.
Common AI image resolutions and their megapixel values
| Resolution | Megapixels | Typical use case |
|---|---|---|
| 512 × 512 | 0.26 MP | Fast drafts, cheap iteration |
| 768 × 768 | 0.59 MP | Standard SD 1.5 |
| 1024 × 1024 | 1.05 MP | SDXL default, DALL-E 3 |
| 1024 × 1792 | 1.84 MP | Portrait / vertical |
| 1792 × 1024 | 1.83 MP | Landscape / horizontal |
| 2048 × 2048 | 4.19 MP | High-res illustration |
| 4096 × 4096 | 16.8 MP | Print-quality outputs |
Note how resolutions that appear to double in size (1024 → 2048) actually quadruple in megapixels and roughly quadruple in compute cost. This squared scaling is the most important thing to understand when budgeting a batch.
Comparing platforms: what the rates actually mean
Stability AI (credit-based). Credits are an abstract unit; the number needed per image scales roughly with resolution tier and model. Mapping credits to dollars requires knowing the current credit pack price, which this estimator uses as an editable default.
Replicate (per-second billing). Replicate charges by the GPU-second for most models. Inference time increases with resolution, so the cost is genuinely tied to computation rather than a tiered pricing structure. This can make Replicate cheaper for low-resolution drafts but more expensive for high-resolution outputs than flat-rate alternatives.
RunPod (per-second billing). Similar structure to Replicate but with different GPU types and pricing. RunPod’s spot instances offer lower rates at the cost of potential preemption, which matters for large batch jobs.
OpenAI Images (per-image tiered pricing). DALL-E 3 charges a flat rate per image at each resolution tier (1024×1024, 1024×1792, 1792×1024). Because cost is fixed within a tier, the per-megapixel figure is an approximation — you pay the same whether your 1024×1024 image is simple or complex.
Tips for keeping batch costs down
- Generate small, upscale later. A common money-saver is to generate at a modest resolution and run a cheaper dedicated upscaler afterwards rather than generating at full size.
- Mind the squared scaling. Doubling both dimensions quadruples megapixels and roughly quadruples cost — resolution choices matter far more than they look.
- Account for retries. Real projects rarely keep the first generation, so budget for 2–4× the nominal image count when planning a large run.
- Confirm minimums. Some platforms have minimum per-call charges or round up to size tiers, so very small images may not be as cheap as the raw per-megapixel math suggests.
- Always confirm current rates. AI image pricing changes frequently. Treat this estimator’s output as a directional budget guide and verify against your platform’s current pricing page before committing to a large production run.